Shallow quantum neural networks (SQNNs) with application to crack identification

Quantum neural networks have been explored in a number of tasks including image recognition. Most of the approaches involve using quantum gates in the neurons. Hybrid neural networks combining classical and quantum layers are recently being studied. The goal of the hybridization is to exploit the generalization benefits of quantum networks while reducing the requisite number of qubits. In this context, a Shallow Quantum Neural Network (SQNN) is proposed in this paper. Such architectures have not been studied previously on image processing tasks. The SQNN is expected to be successful in image classification tasks with limited training set size. Two types of SQNNs have been developed, these are ResNet-SQNNs and VGG16-SQNNs. The SQNN models are applied to the problem of detection of surface cracks on images. Introduction of hybrid classical-quantum layers in a typical pretrained neural network model detects cracks with a greater validation accuracy as compared to classical Res-NNs and VGG16-NNs. Moreover, an entangled feature mapping has been incorporated with the parameterized quantum circuit in SQNNs. This outperforms classical approaches providing improved accuracy and training times. To demonstrate the computational advantage of the quantum neural networks over the classical neural networks, a comparative analysis has been conducted based on efficiencies and improvement in the accuracy achieved for the surface crack detection task.